Gordon’s Signal
“The question is no longer whether we can see a person's risk. We can. The question is what we choose to do with it.”
Key points
- AI can improve the effectiveness and efficiency of underwriting and claims.
- Personalised insight can help insurers support prevention, better health and longevity.
- AI and drones can help organisations anticipate and prevent commercial losses.
- More precise pricing also raises questions about fairness, access and insurability.
Insurance is built on shared uncertainty
Insurance pools losses that are uncertain. We contribute premiums so that those who suffer a loss can receive protection. That shared uncertainty is at the heart of the product.
AI does not remove uncertainty or tell us exactly who will claim and when. But it can make differences in risk more visible, person by person. That is useful for accuracy. It can also put pressure on pooling, pushing more of the cost towards the people most in need of cover.
As more insurers gain access to these capabilities, precision alone will become harder to compete on. Trust will not. So the leadership question is not simply whether to use AI. It is what to do with what AI tells us.
This is not a future debate
The boundaries are already being drawn. The EU AI Act places AI used for risk assessment and pricing in life and health insurance in its high risk category, as EIOPA explains. That is a specific classification, not a blanket statement about every insurance application.
California’s SB 1120 makes human responsibility explicit in medical necessity decisions. Decisions to deny, delay or modify care based on medical necessity must be made by a licensed physician or other appropriately qualified licensed health care professional, not delegated to AI.
Australia has also drawn a line on genetic information. According to the Australian Treasury, its ban on life insurers requesting or using adverse genetic test results in underwriting is scheduled to begin on 8 October 2026.
These rules differ in scope. The leadership lesson is the same: predictive power is not permission. Decide how information will be used before a model makes that choice the default.
What wearable data can reveal about insurance risk
I cofounded MiYA Insurance Solutions, which builds health models from wearable data, so read this section as the view of an interested party. The opportunity is to see how risk changes over time, rather than rely only on a declaration or a single medical examination.
Resting heart rate is one example. A 2017 research review by Aune and colleagues found that each additional ten beats per minute was associated with a 17% higher relative risk of death from any cause. An examination captures a moment. Repeated readings may show a pattern, provided the devices and measurement conditions are reliable.
The direction of travel matters too. In a Norwegian study published in JAMA, people whose resting heart rate rose from below 70 to above 85 beats per minute over about a decade had approximately twice the risk of death from ischaemic heart disease compared with those who remained below 70. This was not a wearable study, but it shows why change over time deserves attention.
Activity can also be measured rather than merely declared. In a 2020 US study, taking 8,000 steps a day was associated with approximately half the mortality risk of taking 4,000. These are observational associations, not proof that changing a reading will produce the same change in an individual’s risk.
The hard work is turning a consumer reading into information fit for an insurance decision. Different devices must produce sufficiently consistent measurements, and clinical or underwriting use requires its own evidence. A wearable is not automatically a substitute for a laboratory test.
Every one of those signals can be used in two ways. To give someone credit for a risk they have reduced, or to charge them more for a risk they carry. The technology is the same in both cases. The choice is ours.
Two kinds of insurance risk, two different answers
Risk that people can change
Activity, smoking and driving behaviour create opportunities to support change after a policy has been underwritten. This is where MiYA sees enormous potential: helping policyholders make healthier choices is good for the individual and can improve the risk of the book over time.
But that opportunity depends on trust. MiYA’s rule is that data shared after underwriting can only improve the policyholder’s position. It may earn them a benefit, but it must never trigger a penalty. Choosing to share information in order to become healthier should not expose someone to higher premiums or worse terms.
That design rule does not settle every fairness question, but it answers the most immediate concern: will the information I share to improve my health later be used against me?
There is evidence that incentives can support sustained activity. In an assessment commissioned by Discovery, RAND Europe found that uptake of Vitality Active Rewards with Apple Watch was associated with about 34% more tracked activity days per month than participation in Vitality Active Rewards alone. The association persisted over the 24-month repayment period. That supports the case for sustained engagement, without proving an equivalent reduction in mortality or claims.
Risk that people cannot readily change
A gene. Age. A condition already diagnosed. A house on a flood plain the owner cannot afford to leave. These are harder cases.
Precision here does not itself reduce the underlying risk. It can move more of the cost onto the person carrying it, including people least able to afford it. Which differences should affect price? Which risks should remain shared?
AI does not remove the need to draw those lines. It makes the decision more urgent. Insurers, governments and communities need to decide how protection remains available when the market’s most precise answer is also its least affordable.
Prevention is not the answer to fairness
I believe in prevention. Anything that helps people live longer and better is good, and insurers are well placed to support it. But be clear about what it solves.
A shared value model needs real engagement and a relationship long enough for the investment to bear fruit. Neither should be assumed. Insurers need to demonstrate that their programmes reach people who need help and produce meaningful benefits.
Prevention is a growth and trust strategy. It is not the answer to who gets charged more and who gets declined. Keep the two arguments apart.
Five decisions for insurance boards using AI
1. Name the factors you will not use
Decide which information should be excluded even when it predicts risk. My view is that genetic results and health data used as a penalty belong on that exclusion list. Set the boundary before a regulator has to set it for you.
2. Test for proxy discrimination
A postcode or a purchasing pattern can act as a substitute for a sensitive characteristic. Test for that possibility, document the method and examine outcomes across customer groups.
3. Make adverse decisions explainable
Customers should be able to understand why they were declined or offered different terms. If your team cannot explain a decline in plain language, you should not be issuing it.
4. Keep a qualified human accountable for claim denials
My proposed rule is simple: no claim is denied without a qualified human reading it. Review must be a meaningful decision, not an automatic confirmation of a model’s recommendation.
5. Own the tail
Where commercial cover becomes difficult to provide, help design a workable pool or shared solution. Leaders should help shape the response rather than quietly withdraw and leave the consequences to others.
Precision is a commodity. Trust is the asset.
My expectation is that increasingly precise risk assessment will become widely available. What will set an insurer apart is whether customers believe their data is used for them, not against them.
That belief is earned by decisions boards take now, in public, before they are forced. Which risks will we still pool? Which data will we use only to give credit? Which decisions must a qualified human sign off?
AI is reshaping how insurers understand risk, support prevention and decide claims. What we choose to do with that knowledge is a leadership choice, not a technology one.
AVER SOLUTIONS PERSPECTIVE
Greater predictive power creates a leadership responsibility: institutions must decide how better information translates into better customer outcomes. Gordon’s challenge to insurers has a parallel in banking. When a bank detects signs of financial difficulty earlier, the response might include timely support, changes to terms or human review. The right response cannot be determined by model accuracy alone.
At AVER Solutions, we believe trust depends on consistency between what an institution promises, what its systems do and how it explains its decisions. Measurable business value should therefore include decision quality, access, customer understanding and outcomes, alongside efficiency. Insurance and banking have different obligations, but both need clear boundaries on how information is used.
ABOUT AVER SIGNALS
AVER Signals, published by AVER Solutions, brings together perspectives from leaders across insurance, banking, technology and transformation, alongside clearly attributed company reflections. Each contribution explores a practical question: as AI reshapes our industry, what should leaders prepare for, and how can we use AI to work smarter, serve customers better and strengthen trust?
The aim is not to predict the future, but to help our industry prepare for it through clear, practical and responsible leadership perspectives.
Questions leaders should consider
How does AI affect insurance risk pooling and customer trust?
AI can reveal differences in individual risk, improving pricing accuracy while putting pressure on shared insurance pools. Gordon Watson argues that leaders must decide which risks remain shared and how customer data will be used. AVER Solutions sees a related banking question: whether an earlier warning of financial difficulty prompts support, tighter terms or both.
Leaders must define the intended customer outcome before automating the response. Insurance and lending have different economics and obligations.
What should insurers check before using wearable data in underwriting?
This is a general underwriting discussion. Gordon’s description of MiYA’s rule concerns information shared after underwriting; it should not be read as a statement about MiYA’s initial underwriting or pricing.
Wearable data needs evidence that it is fit for the intended underwriting decision. Building on Gordon Watson’s distinction between risks people can and cannot readily change, AVER Solutions proposes three questions:
Is it reliable? Check measurement error and performance across relevant customer groups. A population association does not establish suitability for an individual decision.
Can the customer act on it? Identify a realistic improvement, available support and a review point.
What happens to those who cannot? Define the treatment of customers unable to improve or share optional data, with explanations and human review.
Record the evidence, owner and review trigger. This is a discussion framework, not a validated scoring model or a substitute for applicable requirements.
How can insurers and banks turn AI risk signals into customer support?
Define the response before introducing the signal: what support is available, who reviews adverse decisions and how customers can challenge them. These are AVER Solutions’ illustrations, not descriptions of actual programmes.
Insurance: An optional post-underwriting wearable programme would specify reliable readings, eligibility for benefits and how policyholders without usable data are treated. Information shared to support healthier choices should not lead to higher premiums or worse terms.
Banking: An early cash flow warning would trigger checks for a temporary pattern or enduring difficulty, consideration of support, and an accountable review of any changes to terms.
How should insurance and banking boards measure AI customer outcomes?
AVER Solutions recommends a baseline covering:
Decision quality: errors, overrides and successful appeals.
Access and fairness: outcomes across customer groups, including non-participants in optional data programmes.
Understanding: clarity of explanations and complaint outcomes.
Practical benefit: demonstrated health, financial resilience or service improvements.
Review these alongside efficiency and commercial results. Assign owners and correction thresholds before launch. Model accuracy and participation alone do not demonstrate better customer outcomes.
